Impact of Model Order Choice on the Results of Parallel Independent Component Analysis

Parallel independent component analysis (pICA) is a data-driven method that identifies the maximally independent components of multiple imaging modalities while simultaneously investigating the strength of their correlations. Researchers using pICA are given the option to use the suggested model order calculated by the minimum descriptive length (MDL) algorithm, or they can choose their own model order. To date, there are no suggested guidelines for this choice. To test the sensitivity of pICA to the selection of model order, we applied it to a well-researched brain disorder, schizophrenia, looking at the correlations between patterns of grey matter volume (GM) volume and white matter integrity, measured using fractional anisotropy (FA). We varied model orders from low to high, and tested the sensitivity to disorder effects (cases vs controls), similarity of spatial maps identified across model orders, consolidation or distribution effects related to model order selection, and the performance of the minimum descriptive length (MDL) algorithm. The pICA results (multimodal analysis) were also compared to the ICA (unimodal analysis) for each imaging modality. Across model orders, there was consistent sensitivity to disorder effects, and clustered patterns of spatial maps for both the GM and FA reflecting those differences. The MDL-estimated model order captured the majority, but not all, of the spatial patterns present in the GM and FA. There was not the expected consolidation of spatial maps at lower model orders, nor the distribution of spatial maps at higher model orders. The spatial patterns identified in the ICA closely resemble those found in the pICA, although lacking the benefit of the optimization algorithm, were not as highly correlated. This offers some insight and guidance for researchers interested in using pICA with regard to selecting model order for their particular analysis of multiple imaging modalities.